How much better is Gemini Robotics 2.5 than Gemini Robotics 2?

Google DeepMind reports that Gemini Robotics 2.5 has success rates 30% to 100% higher than its baselines, depending on the task, and finishes tasks 20% to 30% faster. In its charts, packing pastries with tongs rises from 5% to 100%, and untying ribbon bows rises from 0% to 70%.

Can developers use Gemini Robotics 2.5 today?

Google DeepMind says Gemini Robotics 2.5 is rolling out to its robotics partners, and the announcement names Apptronik and Mercedes-Benz. The post does not mention general API access, a waitlist or pricing, so most developers cannot run Gemini Robotics 2.5 on their own robots yet.

Which robots does Gemini Robotics 2.5 work on?

Gemini Robotics 2.5 works natively across several robot types, according to Google DeepMind. The post shows Apptronik's Apollo 2 humanoid with its 22-DoF SharpaWave hand, and the Franka Duo platform with two-finger grippers. Apollo 2 ran the Mercedes-Benz kitting pilot.

Does Gemini Robotics 2.5 resist jailbreak prompts?

Google DeepMind reports 99.5% jailbreak resistance for Gemini 4 Argon, against 82% for Gemini Robotics ER 2, on its new ASIMOV-Jailbreak benchmark. The company also says inference-time reasoning guidance raised safe task completion by 33% around obstacles and 43% around nearby people.

Google DeepMind · 2026-11-10 · major

Gemini Robotics 2.5 — DeepMind's most capable robot action model

In short

Gemini Robotics 2.5 is Google DeepMind's most capable vision-language-action model. It was tested on 200+ dexterous tasks, with success rates 30% to 100% higher than Gemini Robotics 2. It reached 96% on a Mercedes-Benz kitting pilot.

Google DeepMind · 2026-11-10 · 4 sources checked

Hero graphic from Google's Gemini 4 Argon for robotics developer post — Google

DeepMind's newest robot model learns fine hand skills like tying bows and serving food with tongs.

Key specs

Mercedes benz kitting success	96%
Speed vs baselines	20%–30% faster

Quick facts

Maker	Google DeepMind
Model type	Vision-language-action (VLA)
What's new	Dexterity: 200+ tasks, tool use, deformable objects
Robots shown	Apptronik Apollo 2, Franka Duo
Pilot	Mercedes-Benz kitting, 96% success
Availability	Rolling out to robotics partners
Open release	ASIMOV-Jailbreak benchmark on Hugging Face

What is it?

Gemini Robotics 2.5 is a vision-language-action model that turns camera input and instructions into robot arm and hand movements. Google DeepMind calls it its most capable such model so far. It was tested on more than 200 tasks, including pipetting liquid, tying a bow on a gift box and handling delicate items.

How does it work?

The action model runs alongside Gemini Robotics ER, an embodied reasoning model that tracks the steps of a workflow. DeepMind also used Gemini 4 Argon as a simulation engine during development. At run time, reasoning guidance steers the robot around physical hazards such as obstacles and nearby people.

Why does it matter?

Hands are the hard part of general-purpose robots, and the reported gains are large on tasks the earlier model failed completely, such as hanging a cardigan (0% to 62%). The 96% result in a Mercedes-Benz kitting pilot shows a real factory-line job, though it is a proof of concept with one partner.

Who is it for?

Robotics teams, humanoid makers and factory automation groups

Frequently asked questions

How much better is Gemini Robotics 2.5 than Gemini Robotics 2?: Google DeepMind reports that Gemini Robotics 2.5 has success rates 30% to 100% higher than its baselines, depending on the task, and finishes tasks 20% to 30% faster. In its charts, packing pastries with tongs rises from 5% to 100%, and untying ribbon bows rises from 0% to 70%.
Can developers use Gemini Robotics 2.5 today?: Google DeepMind says Gemini Robotics 2.5 is rolling out to its robotics partners, and the announcement names Apptronik and Mercedes-Benz. The post does not mention general API access, a waitlist or pricing, so most developers cannot run Gemini Robotics 2.5 on their own robots yet.
Which robots does Gemini Robotics 2.5 work on?: Gemini Robotics 2.5 works natively across several robot types, according to Google DeepMind. The post shows Apptronik's Apollo 2 humanoid with its 22-DoF SharpaWave hand, and the Franka Duo platform with two-finger grippers. Apollo 2 ran the Mercedes-Benz kitting pilot.
Does Gemini Robotics 2.5 resist jailbreak prompts?: Google DeepMind reports 99.5% jailbreak resistance for Gemini 4 Argon, against 82% for Gemini Robotics ER 2, on its new ASIMOV-Jailbreak benchmark. The company also says inference-time reasoning guidance raised safe task completion by 33% around obstacles and 43% around nearby people.

Try it

Read the announcement at deepmind.google; the model is rolling out to robotics partners rather than public APIs.